Examining imaging data of patients from multiple countries, researchers found that a deep learning system demonstrated higher sensitivity and non-inferior specificity for detecting active tuberculosis on chest radiographs in comparison to nine radiologist reviewers.
New research suggests an emerging deep learning system may provide a viable screening resource for detecting active tuberculosis (TB) on chest radiographs.
In a new retrospective study, published in Radiology, researchers compared a deep learning system versus radiologist assessment of chest X-rays of 1,236 patients, 17 percent of whom had active TB. The study authors found that the deep learning system demonstrated higher sensitivity (88 percent vs. 75 percent) and non-inferior specificity (79 percent vs. 84 percent) for detecting active TB in comparison to nine reviewing radiologists.
The deep learning system, which was trained on over 165,700 images drawn from 22,284 patients in multiple countries, also had a receiver operating characteristic (ROC) curve of 89 percent, according to the study.
The study authors said the deep learning system may offer a viable alternative for TB screening in resource-deprived countries.
“The development of a (deep learning system) with high clinical performance across a broad spectrum of patient settings has the potential to equip public health organizations and health care providers with a powerful tool to reduce inequities in efforts to screen and triage TB throughout the world,” explained study co-author Shruthi Prabhakara, Ph.D., an engineering lead with Google Health, and colleagues.
The study authors noted the effectiveness of the deep learning system may allow more targeted use of nucleic acid amplification testing (NATT) to confirm TB.
“Although NAAT techniques have high positive predictive value, the implementation is limited by cost. However, if coupled with an inexpensive but relatively sensitive first-line filter like chest radiography … , NAAT could effectively benefit a larger population due to more targeted use,” noted Prabhakara and colleagues.
The researchers said the deep learning system also maintained comparable effectiveness with radiologists in populations with a higher prevalence of active TB, including HIV-positive patients and a gold mining population test set from South Africa.
“In addition to having a high prevalence of TB, the mining population is also known to have a high prevalence of baseline pulmonary abnormalities like silicosis, emphysema, and chronic obstructive pulmonary disease. We found that (the deep learning system) remained comparable to radiologists in this population,” added Prabhakara and colleagues.
In regard to study limitations, the authors acknowledged the retrospective design of the study and the high prevalence of tuberculosis in the data sets. Accordingly, they noted that the deep learning model will need to be assessed in prospective validation studies as well as populations that have a lower prevalence of TB.